Bentley Moon

Summary

I run a focused, solo research program on a single workstation GPU. Its current core is a competence-gating result: a model's self-correction machinery — verifying another model's output, repairing from execution feedback, resisting poisoned context — helps only inside the region where the model can already tell right from wrong, and can be actively harmful outside it. The law is derived from first principles (false-alarm ≈ 1 − q), confirmed on seven real models (fit −0.01 + 0.83·(1 − q); corr −0.87, replicated ×3), and paired with a second confirmed law: "independent" models share error basins (ρ ≈ 0.55), so ensemble oversight buys less than it appears. The work is deliberately narrow, pre-registered, and honest about scope (local, replicated; frontier validation proposed).

Methodology

• Pre-registered expectations written before each experiment runs.
• Kill criteria enforced — three self-modification experiments killed under the program's own rules.
• Dangerous controls run even when they threaten the thesis (e.g. a 30-line wrapper run against a 2,100-line substrate; result was parity, and the prior overclaim was retracted).
• Three documented overclaim retractions kept in the record, not buried.

Background

Independent and self-directed; no graduate program. Hardware is a single RTX 4090. The through-line across everything is legibility: mechanisms a human can read, audit, and reason about.